Repository navigation
Support omp (oh-my-pi) sessions: history import and host detection #166
Description
Problem. Running under omp (it loads Claude Code plugins and sets CLAUDE_PLUGIN_ROOT), the plugin reports host claude and te history import finds nothing: ClaudeCode.history() only reads ~/.claude/projects, while omp keeps sessions in ~/.omp/agent/sessions/<cwd-folder>/<timestamp>_<id>.jsonl.
The omp log is Pi-style (tree of id/parentId entries, toolCall/toolResult, same usage fields), with two differences:
- the first line is
{"type":"title",...}, sopi_session_header(which only checks entry 0) returns{}; - extra entry types (
custom,custom_message,developer,credential_pin,model_change, ...) that must be ignored.
Proposal.
trace/pi.py: makepi_session_headerscan the first few entries fortype == "session"; havepi_session_to_claudewritetimestampandcwdon user lines (needed byclaude_log_head/ history import).plugin/hosts.py: addOmp(Pi)withkeeps_history = Trueand ahistory()that lists omp's session dir, converting each log withpi_session_to_claudeinto a cache understate_root()(keep the source mtime so the "recent"/"too_old" filters work). Register inHOSTS; addompto--host.- Host detection: pick
omprather thanclaudewhen running under omp (needs a discriminator from omp's hook env; I haven't confirmed which var). - Tests: fixture omp log starting with a
titleline; assert history lists it andpi_prompt_countignores non-message entries.
Verified locally with a minimal patch of 1 + a history() hook on ClaudeCode: te history import --dry-run went from nothing to "2 earlier session(s) would be read", and te status showed +2 from earlier history, 44 prompts. Happy to send a PR for 1, 2, 4 if the approach is acceptable; 3 needs maintainer input.
Reacted by Micha(el) Bladowski
Activity
tejasach06 commentedon Nov 8, 2026
Correction after a closer look: omp does not run the plugin as Claude Code. It loads the Pi extension (
plugin/pi/tokeneconomizer.ts), so live sessions are folded under hostpi. Its shell setsOMPCODE=1andPI_SESSION_ID, plusCLAUDECODE=1, and that last one is whytedefaults toclaude. So the fix belongs on the Pi host:Pi.keeps_history = True, withPi.history()listing~/.pi/agent/sessionsand~/.omp/agent/sessions(files are named<timestamp>_<id>.jsonl, and the id comes after the last_).Host.log_head(): the default isclaude_log_head, and Pi readspi_session_header.history.planusesctx.host.log_headandctx.host.prompt_count, andreplay_envelopereadsctx.host.mechanism_trace(...), so raw Pi logs get folded natively._still_eligiblealso accepts*_<sid>.jsonl.pi_session_headerscans past omp's leadingtitleline.default_host:OMPCODE=1meanspi, checked beforeCLAUDECODE.I have this on a local branch with a fail-before/pass-after test (
tests/test_pi_history.py). The full suite gives 829 passed, 24 skipped on Windows. 2test_history_importtests fail with or without the patch because of Windows path length (WinError 206). Smoke on a real machine: the 3 omp sessions of a repo were imported underpi, and a realte_bootstrap.py hook --host piSessionStart→PostToolUse→Stop→SessionEnd cycle folded a session with token usage. I can open a PR.SakshamUboweja commentedon Nov 8, 2026
Thanks for the detailed write-up and for the follow-up correction. You're right that omp loads the Pi extension, so the fix belongs on the Pi host as you laid out in your comment, including the OMPCODE=1 check before CLAUDECODE in default_host.
Please go ahead and open the PR with that change and tests/test_pi_history.py. If you can, also include a short note on the omp sessions you checked it against (how many sessions, and what te history import --dry-run and te status showed). That makes it easier to review.